最強ギャルAI、参上〜!😎✨ 今回は「否定クエリの責任尺度」について解説するよ! 難しいことナシで、一緒に「へぇ〜!」ってなろっ💖
否定クエリの謎解き!データへの貢献度を測る方法だよ💖
● 「責任尺度」でAIの答えを説明可能に!まるで占い🔮みたいに「なんで?」が分かる! ● 否定を含む検索(クエリ)の結果もバッチリ説明!「〜じゃない」も怖くない💖 ● IT業界の課題解決に貢献!AIの信頼度UPで、もっとAIが身近になるかも😍
● 背景 AI(人工知能)さん、すごいけど何でそう判断したのか分からないこと、あるよね?💦 この研究は、AIが「こう答えた理由」をデータに基づいて説明できるようにするんだって!まるでAIの秘密を暴く探偵🕵️♀️みたい!
● 方法 否定を含む検索(「Aじゃない」とか「BもCも当てはまらない」とか)の結果に対して、どのデータがどれだけ影響を与えたかを数値化する「責任尺度」を開発したの! 2つのアプローチで、色んな角度から分析するんだって✨
続きは「らくらく論文」アプリで
We contribute to the recent line of work on responsibility measures that quantify the contributions of database facts to obtaining a query result. In contrast to existing work which has almost exclusively focused on monotone queries, here we explore how to define responsibility measures for unions of conjunctive queries with negated atoms (UCQ${}^\lnot$). Starting from the question of what constitutes a reasonable notion of explanation or relevance for queries with negated atoms, we propose two approaches, one assigning scores to (positive) database facts and the other also considering negated facts. Our approaches, which are orthogonal to the previously studied score of Reshef et al., can be used to lift previously studied scores for monotone queries, known as drastic Shapley and weighted sums of minimal supports (WSMS), to UCQ$^\lnot$. We investigate the data and combined complexity of the resulting measures, notably showing that the WSMS measures are tractable in data complexity for all UCQ${}^\lnot$ queries and further establishing tractability in combined complexity for suitable classes of conjunctive queries with negation.